Web capture

State: to-be-approved. Deploy-verified against a live workspace: not yet. Treat Done when below as the acceptance test and review cargo-ai project plan before deploying. Make no outcome claim for this skill until it is approved.

The outcome

What the company and its competitors say on the web, and what the web says about them, stops drifting away from what every agent reads. Once a week one agent runs, and lands it in your repository:

1. The raw reads. One parallel.extract call reads the company's own pages (home, pricing, customers, careers, blog, changelog) and each competitor's watched pages (usually pricing and changelog), in full. One parallel.createTask news question covers all of them for the days since the last merged read. A fixed node -e line writes each page to its own file under cadence/log/raw/web/ and the news to cadence/log/raw/web/news/<date>.json; no text passes through the agent. git diff against the default branch is what says a page changed.
2. The context files. The first run seeds every empty domain: global/ (positioning, value proposition, offerings), icp/ as an inferred profile with at least one disqualifier, alternative/, client/, proof/ and signal/ candidates. Every run after adds one dated insight/<date>-web.md with what changed that week, for the company and each competitor, plus a client/ or alternative/ file for a newly named customer or competitor. It never edits a file that exists: a change to one is a proposal in the pull request body.

Then it opens one pull request and stops. A human merges, and the next cargo-ai project deploy syncs context/ into the workspace context repository, which is where every other Cargo agent reads before it acts. A week with nothing worth writing opens no pull request.

Every factual sentence carries an evidence tag: [R: <url>] receipted by a page, [I: <from what>] inferred, [TR: <what would settle it>] unknown. Public pages are one source, so nothing here is stated with the conviction of a deal; win-loss-review is what later verifies the ICP against won and lost.

The whole cookbook is one agent and its prompt. There is no collector script: the prompt spells out the exact commands, generated from the domain, the pages and the competitors at the top of the prompt file, and git does the diffing. Three properties make it safe enough to run unattended:

• The reads are fixed and mechanical. The same two commands every week, the full page content (not excerpts, which change with the question asked), and files written by code rather than retyped by the agent. Two reads a minute apart came back byte-identical, so a diff is a change on the site. The news window starts at the last merged read, so a quiet week is covered by the next run rather than skipped.
• The pull request is the gate. The agent has repository write access and nothing else: no workspace context write, no CRM, no email, no Slack, no merge.
• Append-only after the first run. Existing files are never edited, so what a human corrected stays corrected.

Example

Every Monday, read our website, our competitors' pricing and changelog pages, and the news about all of us, and keep context/ current.

Illustrative output, fictional records:

insight/2026-10-05-web.md
- Northwind: new "Enterprise" tier on /pricing, with SSO and audit logs [R: https://northwind.example/pricing]
- Northwind: raised a $20M Series B led by Contoso Ventures, 2026-10-01 [R: https://news.example/northwind-series-b]
- Northwind: Fabrikam named as a customer, 40% faster onboarding [R: https://northwind.example/customers/fabrikam]
- Tailspin (competitor): Team plan price up from $49 to $69 a seat [R: https://tailspin.example/pricing]

client/fabrikam.md, proof/fabrikam-onboarding.md   (reference_permission: unknown)
PR body: proposes adding "Enterprise" to global/offerings.md and the new price to alternative/tailspin.md

A week after the first run: one pull request, four findings across two companies, two new files, two proposed edits, and two reads billed (the pages and the news).

Put it in your project

This folder is a worked example: real project resources written for some other company. The job is to end up with the code your company would have written, in your project, and an agent does the adapting. If the cargo-project skill is in your session it carries the long form of this; if not, this is enough.

1. Install it: the CLI does the copy. From inside the project (the folder holding cargo.json), cargo-ai project add cookbook/web-capture writes the resources to infra/web-capture/ and this procedure to .agents/skills/web-capture/. No project yet? cargo-ai project init <dir> --cookbook web-capture && cd <dir> && npm install does both; this folder never ships a shell. If you are reading this from the project's .claude/skills/, the install already happened: start at step 2. On a CLI too old to have add, copy this folder in as a sibling of what is there by hand; everything below is unchanged.
2. Reconcile it with what is already declared. If the project already has a GitHub or Anthropic connector or an agents folder, rewire the imports to the existing one and drop the copy; two resources with one slug is a collision at deploy. The knowledge layer needs no work: the scaffold already declares the repo's root context/ in infra/context.ts, and defineContext is a per-workspace singleton, which is why this folder ships none. Append nothing to .env.example: nothing here holds a credential.
3. Set the domain, the pages and the competitors. In infra/agents/web-scribe.prompt.ts (at infra/web-capture/agents/web-scribe.prompt.ts once installed), set DOMAIN to the company's own domain, PAGES to the site's real sections (the sitemap lists them), and COMPETITORS to the competitors worth watching with their pricing and changelog URLs. The agent refuses to run on the placeholder domain; with no competitors it watches the company alone and says so in every pull request.
4. Adapt. Work the sections below in order: What should not change is what you argue back about (say what breaks, then do it if they still want it); What you can change is what you offer unprompted; What you will be asked is the floor, and you derive before you ask. Set the model. Record what you changed and why under a ## Decisions section in your copy of this file.
5. Plan, then stop. node --import tsx evals/contract.mjs from this skill's folder, then npm run check && cargo-ai project plan from the project root (check validates the resource tree offline; the blank template ships it). Show the diff: one agent, two bound connectors, one folder, and no model. Deploy only on an explicit yes: cargo-ai project deploy. Never cargo-ai project init --force into a non-empty directory.
6. Run the first one by hand. From the workspace UI or with cargo-ai ai message create --agent-uuid <uuid> --parts '[{"type":"text","text":"Run the web capture. Follow your system prompt exactly."}]'. The first run always opens a pull request: its page and news files are the baseline every later week is diffed against, and they only count once merged. The Monday cron takes it from there.
7. Verify. Walk Done when line by line and report each with evidence.

What you will be asked

Derive before you ask. An input with a lookup is looked up, not asked. Only the rows marked asked genuinely live in the operator's head.

Checked before moving on, not after the deploy:

• DOMAIN is the company's own domain, and PAGES and COMPETITORS name pages that exist
• the contract passes, which includes the 20-URL limit
• cargo-ai project check prints the agent bound to the repository root, not infra/
• exactly one defineContext in the project, the scaffold's, resolving to the root context/

What you can change

The code is a worked example. These reshapes are expected, and the agent offers them rather than waiting to be asked. Every one costs something; that is what makes it a variation and not the default.

What should not change

However far you adapt, these hold. Ask for one anyway and the agent tells you what breaks, then does it if you still want it, and records why under ## Decisions in your copy of this file.

• The web is the only source. (web-scribe.prompt.ts) No CRM, no calls, no inbox, no Slack. Each is another cookbook's evidence; mix one in here and a page and a deal end up in one file at one confidence nobody can tell apart afterwards.
• The reads are the commands in the prompt, with fullContent, and code writes the files. (web-scribe.prompt.ts) Excerpts change with the question asked, and an agent that retypes a page truncates and tidies it; either makes every week look like a change on the site.
• outputSchema stays a string. (web-scribe.prompt.ts) The news action takes the JSON schema as a string; an object fails its validation and the week reads no news.
• Existing files are never edited. (web-scribe.prompt.ts) What a human corrected stays corrected. A change to an existing file is a proposal in the pull request body.
• The one-PR write path. (infra/agents/web-scribe.ts) The agent never writes the workspace context directly and has no context capability: the repository is the source and cargo-ai project deploy syncs it.
• The first run always opens a pull request. (web-scribe.prompt.ts) Its page and news files are the baseline. Until they are merged, every run reads as a first run and nothing is ever a change.
• Every claim is tagged. (web-scribe.prompt.ts) [R], [I] or [TR], and missing evidence is never contradicting evidence.

Done when

• DOMAIN, PAGES and COMPETITORS name the company, its real sections and the competitors worth watching, and the contract passes
• cargo-ai project check prints the agent bound to the repository root, and cargo-ai project plan reports one agent, two bound connectors, one folder and no model
• the first run opened one pull request with the page, competitor and news files, seeded every empty domain, and edited no existing file
• every seeded file carries tags, icp/ names a disqualifier, and every client/ file carries reference_permission: unknown
• after the baseline was merged, a run with a change opened a pull request adding insight/<date>-web.md with one tagged line per finding, each naming its company, and proposed rather than made any edit to an existing file
• a run with nothing worth writing opened no pull request
• after merging and cargo-ai project deploy, a seeded file is readable from the workspace context repository and an agent with the context capability quotes it back with its tag

What it costs

Fetch live prices before every estimate: cargo-ai connection integration get parallel, plus the LLM connector's model. Quote the lookup time with the estimate.

Each run bills two reads: one parallel.extract over every listed URL, the company's and the competitors' (billed per URL, a 404 included), and one parallel.createTask on the lite processor, the cheapest rung of its price ladder. Plus the harness run, which scales with how much changed.

Composes into

tam-building and account-scoring (they read the ICP this writes down: one to source the market, the other to score the book), new-hire-detection (it qualifies the company each new hire joined against that ICP), website-building (the positioning this captures is what the pages say), win-loss-review (verifies the ICP this seeds against won and lost deals, and proposes the corrections), and any agent with the context capability.